Model Configuration

Note: to access the model configuration files, users need to clone the code repository (installation via pip is insufficient).

During query processing, ThalamusDB dynamically selects the most suitable language model for each semantic operator. Currently, only models by OpenAI are supported. Users can configure the way in which ThalamusDB selects models. The associated configuration file is located at config/models.json. This is an extract of the default version of the file:

{
	"models":[
		{
			"modalities": ["text", "image"], "priority": 10,
			"kwargs": {
				"filter": {
					"model": "gpt-5-mini",
					"reasoning_effort": "minimal"
				},
				"join": {
					"model": "gpt-5-mini",
					"reasoning_effort": "minimal"
				}
			}
		},
		...
	]
}

Each entry in the models list is a dictionary with the following properties:

Property	Semantics
modalities	A list of supported data modalities
priority	ThalamusDB prefers models with higher priority
kwargs	The configuration to use for each semantic operator

The following data modalities are recognized:

• text
• image
• audio

When selecting models, ThalamusDB first filters to the models that support all required data types. Note that ThalamusDB supports joins across different data modalities (e.g., matching images with associated text descriptions). In that case, ThalamusDB requires models that support all relevant data modalities.

After narrowing down the choice to the models that support all required data modalities, ThalamusDB considers the priority. Among all eligible models, ThalamusDB selects a model with the highest priority. Ties are broken arbitrarily.

The kwargs field contains the keyword parameters to submit for model calls, separated according to the two semantic operators currently supported by ThalamusDB (i.e., join and filter). At a minimum, the set of parameters must include the model parameter, specifying the ID of the model to use (e.g., gpt-5-mini). Internally, ThalamusDB uses the LiteLLM framework to call language models. Therefore, any parameter that can be used with this framework is admissible and is directly passed on to the completion function.